AI Nutrition & Ingredient Intelligence: Built From Care, Guided by Evidence
Architectural Challenge & Business Bottleneck
Built to help my wife better understand how meals, ingredient quality, preparation, sleep, physical activity, and daily routines relate to how she feels. Everyday observations rarely arrive as clean rows, making weeks of overlapping habits hard to summarize for healthcare conversations.
Production Stack
Engineered Solution & Data Pipeline Topology
- ▹Built multimodal conversational capture across Telegram (text, voice memos, plate photos, barcodes) and React 19 with a 3-stage Gemini Gatekeeper that asks one concise follow-up when missing prep or timing context matters (0 partial rows written in tests).
- ▹Paired FatSecret OAuth 2.0 lookups and deterministic FDA/Monash additive rules (polyols, gums, emulsifiers) with shared Zod contracts and tenant-partitioned dbt Bronze -> Silver -> Gold 1 PRECEDING lag windows (410/410 synthetic SQLite parity).
- ▹Combined exploratory BigQuery ML (XGBoost and ARIMA_PLUS) chronological holdouts, modeled baseline intervals, abs(r) >= 0.35 collinearity warnings, and Semantic Scholar MeSH literature references to surface honest questions for clinician review.